Planning and exploration in stochastic relational worlds
Tobias Lang · Universitätsbibliothek der FU Berlin Hochschulschriftenstelle u. Dokumentenserver · 2011
Goal-directed behavior is one of the most interesting aspects of human and animal intelligence. This thesis addresses planning and exploration in so called stochastic relational worlds which are characterized by two key attributes: they contain large numbers of objects whose properties and relationships can be manipulated, and the effects of actions are uncertain. Such worlds comprise many natural environments such as households, offices, or factories. We take up ideas from the emerging field of statistical relational artificial intelligence and combine rich symbolic representations with a probabilistic framework to learn and represent compact models of action effects which generalize across objects. We propose a variety of methods for planning with such models in ground relational domains on the level of concrete objects. Our methods are largely based on the information processing principle of probabilistic inference in graphical models. We introduce a framework for focusing on relevant objects in planning. We lift existing exploration theories to relational representations. This results in a novel form of exploration which focuses decidedly on objects to which the learned knowledge does not generalize yet. Combining our proposed techniques with existing methods enables goal-directed behavior of autonomous agents in stochastic relational worlds.